Keras（TensorFlow后端）：层输出跨层输入的模型搭建报错求助

Hey there! Let's work through the errors you're seeing with your Keras model implementation—especially that "Layer conv1 has no inbound nodes" message—and get your non-sequential model on track.

First, Let's Identify the Key Problems #

• Duplicate Layer Names: You've named all three Convolution3D layers conv1—Keras requires unique names for every layer, so this is causing immediate conflicts.
• Missing Input Tensor Connection: You're trying to access conv1.output without first connecting the layer to an explicit input tensor. For functional API models (which you need for non-sequential structures), you always start with an Input layer that defines your input shape.
• Unstructured Functional API Usage: You're creating layers but not properly chaining them together or wrapping them in a Model instance, which Keras needs to track the full computational graph.

Corrected Code Example #

Here's how to rewrite your code to fix these issues, following Keras' functional API best practices:

from keras.layers import Input, Convolution3D, ConvLSTM2D
from keras.models import Model

# Define your input shape parameters (replace with your actual values)
patch_size = 10  # Example value
img_rows = 64    # Example value
img_cols = 64    # Example value

# Step 1: Create the core input tensor
input_tensor = Input(shape=(patch_size, img_rows, img_cols, 3), name='input_layer')

# Step 2: Build the first branch (X1 -> Y1 and X2)
conv1 = Convolution3D(32, (3, 3, 3), activation='relu', padding='same', name='conv1')(input_tensor)
lstm1 = ConvLSTM2D(filters=3, kernel_size=(3, 3), padding='same', return_sequences=True, name='lstm1')(conv1)

# Step 3: Build the second branch (X2 -> Y2 and X3)
conv2 = Convolution3D(32, (3, 3, 3), activation='relu', padding='same', name='conv2')(conv1)
lstm2 = ConvLSTM2D(filters=3, kernel_size=(3, 3), padding='same', return_sequences=True, name='lstm2')(conv2)

# Step 4: Build the third branch (X3 -> Y3)
conv3 = Convolution3D(32, (3, 3, 3), activation='relu', padding='same', name='conv3')(conv2)
lstm3 = ConvLSTM2D(filters=3, kernel_size=(3, 3), padding='same', return_sequences=True, name='lstm3')(conv3)

# Step 5: Define the model (specify inputs and outputs; adjust outputs based on your merge needs)
# For example, if you want to output all three LSTM results:
model = Model(inputs=input_tensor, outputs=[lstm1, lstm2, lstm3])

# Optional: Print the model summary to verify the structure
model.summary()

Key Fixes Explained #

1. Unique Layer Names: Each Convolution3D and ConvLSTM2D now has a unique identifier (conv1, conv2, conv3 and lstm1, lstm2, lstm3), eliminating naming conflicts.
2. Explicit Input Layer: We start with an Input tensor that defines the input shape, which is passed directly to conv1—this gives conv1 its required inbound node, fixing that error.
3. Proper Chaining: Each layer is connected to the previous tensor (e.g., conv2 takes conv1's output as input, matching your model's structure where X1 feeds into X2).
4. Model Wrapping: The Model class wraps the input and output tensors, so Keras can properly track the entire computational graph.

Once you've got this base structure working, you can add your merge logic for the LSTM layers (using layers like Concatenate, Add, etc.) depending on how your model needs to combine those outputs.

内容的提问来源于stack exchange，提问作者Mohamad Ballout